Projektbeschreibung
Aufdeckung des sensorischen Prozesses von Ruderfußkrebsen
Winzige Krebstiere, die als Ruderfußkrebse bekannt sind, sind trotz Meeresströmungen und Turbulenzen in Meereslebensräumen im Überfluss vorhanden. Ihre Antennen reagieren empfindlich auf mechanische und chemische Informationen, die es ihnen ermöglichen, Nahrung, Raubtiere und Gefährten zu lokalisieren. Das EU-finanzierte Projekt C0PEP0D zielt darauf ab, mithilfe von Anwendungen des Verstärkungslernens zu bestimmen, wie die Ruderfußkrebse mechano- und chemosensorische Informationen in einer turbulenten Meeresumwelt verarbeiten. Dazu wird ein virtuelles Lernmodell mit hydrodynamischer Signalübertragung erstellt, wo mit lebenden Proben in simulierten Umgebungen experimentiert wird und neuartige Algorithmen zum Verstärkungslernen entwickelt werden, welche die Evolution und das Lernen von Ruderfußkrebsen nachahmen. Letztendlich wird das Projekt einen Einblick in die Entwicklung von Meeresarten geben und könnte zukünftige Innovationen in der biomimetischen Technik inspirieren.
Ziel
Life is tough for planktonic copepods, constantly washed by turbulent flows. Yet, these millimetric crustaceans dominate the oceans in numbers. What have made them so successful? Copepod antennae are covered with hydrodynamic and chemical sensing hairs that allow copepods to detect preys, predators and mates, although they are blind. How do copepods process this sensing information? How do they extract a meaningful signal from turbulence noise? Today, we do not know.
C0PEP0D hypothesises that reinforcement learning tools can decipher how copepod process hydrodynamic and chemical sensing. Copepods face a problem similar to speech recognition or object detection, two common applications of reinforcement learning. However, copepods only have 1000 neurons, much less than in most artificial neural networks. To approach the simple brain of copepods, we will use Darwinian evolution together with reinforcement learning, with the goal of finding minimal neural networks able to learn.
If we are to build a learning virtual copepod, challenging problems are ahead: we need fast methods to simulate turbulence and animal-flow interactions, new models of hydrodynamic signalling at finite Reynolds number, innovative reinforcement learning algorithms that embrace evolution and experiments with real copepods in turbulence. With these theoretical, numerical and experimental tools, we will address three questions:
Q1: Mating. How do male copepods follow the pheromone trail left by females?
Q2: Finding. How do copepods use hydrodynamic signals to ‘see’?
Q3: Feeding. What are the best feeding strategies in turbulent flow?
C0PEP0D will decipher how copepods process sensing information, but not only that. Because evolution is explicitly considered, it will offer a new perspective on marine ecology and evolution that could inspire artificial sensors. The evolutionary approach of reinforcement learning also offers a promising tool to tackle complex problems in biology and engineering.
Wissenschaftliches Gebiet
- natural sciencesbiological sciencesecology
- natural sciencescomputer and information sciencesartificial intelligencemachine learningreinforcement learning
- engineering and technologyelectrical engineering, electronic engineering, information engineeringelectronic engineeringsensors
- natural sciencescomputer and information sciencesartificial intelligencecomputational intelligence
Programm/Programme
Thema/Themen
Finanzierungsplan
ERC-ADG - Advanced GrantGastgebende Einrichtung
13383 Marseille Cedex 13
Frankreich